Running an assistant fully on your machine avoids sending nearby speech, diaries and personal context to the cloud — but it only works when the app manages model choices, memory and device limits without burdening the user. Jarvis aims to be that desktop companion: a voice‑first assistant that keeps transcripts and long‑term memory locally, routes speech and chat through a local model server of your choice, and redacts sensitive items before saving.
What Sets It Apart
- Local-first model flexibility — Works with Ollama or any OpenAI-compatible endpoint (LM Studio, llama.cpp, vLLM, etc.), and surfaces model memory/VRAM budgets so you can pick models that fit your hardware. So what: you control where inference runs and can balance latency vs capability without shipping data to third-party services.
- Voice-first, conversational context — Maintains a short rolling transcript of nearby speech so Jarvis can join an ongoing conversation and follow-up naturally. So what: it treats spoken interactions as continuous context, not isolated commands, improving follow-ups in multi-speaker settings.
- Personal memory and integrations — Stores a local diary and knowledge graph, supports dictation into other apps, and integrates external tools via the MCP protocol (browser automation, smart-home, GitHub, etc.). So what: you get context-aware automation and searchable history without cloud storage.
- Privacy and boundary controls — Sensitive fields are redacted before model context or saved diary entries; network features (search, location, MCPs) can be disabled to reduce external access. So what: you can minimize outbound data while retaining useful functionality.
Who It's For and Tradeoffs
Great fit if you want a desktop voice assistant that keeps data local, can run on your own model server, and supports dictation, TTS/STT and tool integrations. Jarvis is useful for privacy-minded users, developers building desktop AI workflows, and people who want a companion that can listen briefly and join conversations.
Look elsewhere if you need a mobile-first product (no mobile app), turnkey cloud-hosted accuracy without local model management, or a guaranteed cross‑platform feature parity (Jarvis is developed primarily on macOS and some features differ on Windows/Linux). Also note dictation has platform limitations (e.g., unavailable on macOS 26+ in current versions) and very large models require sufficient VRAM or a remote endpoint.
